Skip to main content

Using machine learning in a completely new way to improve climate models

Research facilitated by LIDA has created a new way to understand one of the most uncertain parts of our current climate models. Using unsupervised machine learning in a never-been–done-before approach, the work is the first unsupervised neural network model which autonomously discovers cloud organization regimes in satellite images. This has vitally important implications for weather predictions and climate models.

Climate model calculations disagree by several degrees, and the representation of shallow trade-wind convective clouds dominates the uncertainty in these, so there’s an urgent need to better understand how these clouds form and interact with their environment. “We don’t fully understand the radiative impact of cloud spatial organisation,”  said Dr Leif Denby, author of this new research. “The way these specific clouds organise into patterns affects the radiation balance of heat from the sun. Because of their structure they are cooling the earth – so they are incredibly important. Current climate models don’t agree on whether these clouds will decrease in amount, which has impacts on our ability to understand and model climate change.”

His study of tropical ocean clouds ( – in review) led to Dr Denby’s involvement in the major international research campaign, the EUREC4A (Elucidating the Role of Clouds-Circulation Coupling in Climate). As part of this, he, along with a team of UK researchers, flew and measured atmosphere and clouds in the Caribbean for a month with the aim of gaining a better understanding of the behaviour of clouds and their role in climate change.

The UK part of the project (EUREC4A-UK), is led by researchers at the University of Leeds and National Centre for Atmospheric Science (NCAS), plus external partners including the British Antarctic Survey (BAS), the Met Office and the University of Manchester. The team made detailed measurements showing how shallow trade-wind clouds form and evolve.

“As humans we see cloud structure but it is incredibly hard to measure them. I thought, with all the machine learning we have these days, a machine could look at all of these images and work out the patterns,” explains Dr Leif Denby.

The difficulty was, this type of machine learning was something that had never been done before. Dr Denby was inspired by an unsupervised learning technique used by Neal Jean and colleagues at Stanford University (N. Jean et al 2018, in their research “Tile2Vec: Unsupervised representation learning for spatially distributed data”. An unsupervised technique is where the model is not informed what the correct answer is (as would be the case for supervised learning) but rather the similarity in answers between different inputs is used to constrain the model’s behaviour. In Neal Jean’s work, researchers inputted 3 images simultaneously, and instead of training the neural network to produce three specific answers, the network was encouraged to produce similar answers for two of the images. These two images (in each training example) were selected from a nearby region, and therefore on average contain similar clouds, whereas the third image was sampled at a random location. Through this specialised spatial sampling and constrained learning objective, the machine learned to group things via similarity even though it didn’t know what ‘similar’ was.

This ground-breaking research used this machine learning technique to understand and identify different types of cloud formation. The research employed a neural network, a computing system inspired by biological neural networks in animal brains where an interconnected group of nodes can transmit to one another (like synapses in the human brain) and extract spatial structure in images. These neural networks can learn, or be trained by input and results. 10,000 images constructed from the NASA/NOAA GOES-16 geostationary satellite observations were inputted and from that the system was able to automatically discover different forms of cloud organisation. From this it could learn to group images containing similar cloud structures together. The clusters of similar cloud structures that were found have distinct radiative properties and these properties can now be more fully understood.

Understanding clouds and how they form is something that researchers have struggled with. It’s impossible to model every single meter in the atmosphere to study how clouds develop, but studying the clouds by applying machine learning directly to the vast amounts of satellite data available provides a unique opportunity to understand cloud formation. This research is a way to understand and quantify cloud behaviour that we have never had before. It will make important contributions to work on weather predictions and climate change. We can apply this tool to models and observations, for example identifying 100km meso-scale convective storms that travel over West Africa and become tropical cyclones.

But the work has far-reaching effects outside of weather – it can be applied to any kind of geophysical image. By using an unsupervised approach, there is no need to hand label the training data, and the neural network doesn’t just recognise predefined classes of organisation. It maps similarity and transitions, and can be applied to any spatial dataset with limited effort, so it enables important research in environment and beyond.

More information:

For the full research paper:

For more information on the EUREC4A- UK project:

For the international EUREC4A project:

Back to annual review